Agentao: A Governed Local-First Runtime for Tool-Using LLM Agents
A recent study released on arXiv introduces Agentao, a novel runtime system intended for the management of local-first large language model (LLM) agents using diverse tools. This system seeks to mitigate risks such as excessive permissions and inadequate auditing, as well as concerns about prompt injection and tool contamination. Agentao features a layered architecture that separates model suggestions from user-approved actions and incorporates interfaces for hosts and a contract system. Additionally, it includes a permission-oriented tool structure and subsystems for memory, plugins, skills, and more, focusing on governance and auditing to ensure secure operation of AI agents.
Key facts
- Paper arXiv:2608.13574 introduces Agentao, a governed local-first runtime for tool-using LLM agents.
- Agentao addresses risks including over-privileged actions, weak auditability, prompt injection, tool poisoning, and uncontrolled side effects.
- The system separates model-generated action proposals from host-authorized execution.
- Architecture includes host-facing surfaces, a host contract, a runtime core, a permission-mediated tool system, and supporting subsystems for memory, replay, plugins, skills, sub-agents, and protocol integration.
- The paper covers motivation, threat model, design goals, governance model, execution pipeline, and structured event interface.
- Agentao does not provide formal safety guarantees.
- The paper is published on arXiv with announcement type 'new'.
Entities
Institutions
- arXiv